mosaicml / llm-foundry

LLM training code for Databricks foundation models
https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm
Apache License 2.0
3.84k stars 503 forks source link

Remove spurious warning #1258

Closed dakinggg closed 1 month ago

dakinggg commented 1 month ago

This warning isn't helpful (and is confusing for the finetuning case). Only text dataloader accepts eos/bos token id, to be used for per sequence attention masking on pretokenized and concatenated sequences. Essentially all tokenizers have an eos/bos, and if you use mpt with attn_uses_sequence_id (to enable per sequence masking) without an eos/bos specified, you will get an error separately from this one. finetuning dataloader does not accept this.